How the Trust Index measures AI recommendation share

1. What the index measures

The Trust Index reports which eligible businesses AI assistants named in response to standardized prompts during a stated measurement window. The metric is AI recommendation share: the share of prompt × engine × run cells in which an assistant named the business, weighted by position in the answer.

It measures assistant visibility. It does not measure business quality, safety, regulatory standing, customer outcomes or consumer behaviour. A high share means assistants name the business often; a zero share means no assistant named it in the window. Neither is a judgment about the business.

Two things this index does not claim:

• That consumers act on assistant recommendations. Assistant output and buyer behaviour are separate evidence; the index publishes only the first.
• That assistants ranked anything. The publisher ran the prompts and computed the metric. Assistants answered questions.

Every reading is dated, versioned to this methodology, and retained. Readings are never overwritten.

2. Market universe

Every ranking starts from a published list of eligible businesses drawn from a regulator or licensing body, not from what assistants happen to name. A vertical without an authoritative list is published as an editorial pilot with a constructed universe and labelled as such.

For each vertical × geography the index publishes: the source list and its date, the number of eligible businesses, the number named by at least one assistant, exclusions and the reason, and the duplicate-resolution rule (brands, professional corporations and shared firm names resolved to one entity).

Duplicate resolution uses the regulator's own grouping where one exists: HCRA assigns every licensee to an umbrella entity, so a builder's project companies (for example, 45 licensees under Mattamy) count as one business. Where a regulator publishes no grouping, the index resolves brands, professional corporations and shared firm names to one entity and publishes the rule.

Activity threshold. Where the regulator publishes activity data, eligibility requires a minimum level of it, so the universe is made of businesses a buyer could actually hire: for home builders, at least 100 Tarion possessions in the last 10 years across the entity's licensees (about ten homes a year). Businesses below the threshold that assistants name anyway are listed as named, not eligible and are not scored. The threshold and the counts above and below it are published with every reading.

Exclusions by type. Some licensees are not the business a buyer chooses. For home builders, construction managers and general contractors (firms licensed because they build under contract for a developer or owner, not because they sell homes under their own brand) are a separate category: they are excluded from the builder universe, counted, and not named in builder rankings. For immigration consultants, universities, colleges, hospitals and other employers whose licensed advisers serve their own students or staff are excluded, and so are law and legal practices, which answer to a law society rather than the College.

Where size is a poor proxy. Some registers publish no activity data, only who is licensed. For immigration consultants the base list uses the number of licensed consultants per firm (at least 2), but a well-known practice is often one licensed consultant with a support team. So in that vertical, a firm named by an assistant during a capture is matched to the register by firm name, trade name or consultant name; if at least one of its consultants is entitled to practise, it is added to the index at that reading and scored, labelled with the assistant, date and question that surfaced it. Base-list firms that no assistant names are reported as not named. The base list alone defines that report.

Named, not matched. A business an assistant names that cannot be matched to the register is reported in a separate table with what was checked (for immigration: the CICC register by firm and consultant name, provincial law societies and the Chambre des notaires du Québec). Not finding a match is not a finding that the business is unlicensed: it may practise through a lawyer or notary, trade under a name the register does not show, or be a referral or information site. The table states the checks and draws no conclusion.

Outside the rule above, a business is not added because it was named, and not removed because it was not. Minimum eligible-universe size for publication: 20.

3. Prompt design

Each vertical uses a fixed set of 17 prompts (16 scored questions and one branded question), published in full, written the way a buyer phrases the question inside an assistant rather than the way the industry phrases it on a search engine. The set does not change between readings; a change creates a new methodology version.

The 16 scored prompts sit on five pillars. Every pillar carries at least two prompts. Pillars are not weighted: every scored prompt counts equally, so a pillar's influence on the score is its number of prompts.

Branded prompts (“Is [business] a good company to buy from?”) are run for every eligible business and reported on its entity page. They are excluded from the ranking score, because a branded prompt measures recall of a name the user supplied, not authority.

Prompts are written in plain language, name no business, and are reviewed once a year against how buyers phrase the question in forums and support channels.

4. Capture protocol

A reading is one 72-hour window, monthly, in which every prompt is put to every engine several times under identical settings.

Prompts are submitted to the assistants' consumer products through a third-party capture service (SearchAtlas Prompt Simulator), with the country set to Canada and no account history. The service returns the full answer, every business named with its position, and every source cited. The publisher does not type prompts by hand and does not use the assistants' developer APIs, whose answers differ from what a buyer sees.

For every cell the index preserves: prompt text, engine, model identifier where shown, account type, location setting, timestamp, full response text, every business named and its position, every source URL cited. Raw responses are retained internally for the life of the index and are not published by default (section 6).

AI answers are non-deterministic. Repeated runs exist to measure that, not to hide it; the spread across runs is published for every business (section 5). A reading in which an engine fails or changes behaviour mid-window is noted on the reading and the affected cells are marked.

5. Scoring

A business's recommendation share is the position-weighted sum of the cells naming it, divided by the total number of cells in the reading, scaled to 0–100. A reader with the published cell file can recompute every score and rank.

where N is the number of unbranded prompt × engine × run cells (16 questions × each scored assistant × 3 runs; each reading lists the assistants it scored and its answer count), and w(p) is the weight for the position p at which business b appears in cell c (w = 0 when not named). Position is the order in which businesses are named in the answer.

Position weights (since v2.2). w(p) = 1 / log₂(p + 1), the discount used to score ranked lists in information-retrieval evaluation (discounted cumulative gain). It is defined for every position, favours the first name without making later names worthless, and sits between observed search click-through curves (steeper) and a flat count (no discount).

A business named first by every engine on every scored prompt in every run scores 100. A business never named scores 0.

Published per business: share, spread (standard deviation of share across the three runs), rank, prompts named in (count of the 16 scored questions), engines naming (count of the scored engines), first positions, sources cited, the result per engine, and change since the prior reading.

Rank orders businesses by share. Ties share a rank. Rank is reported against the eligible universe (“#1 of 54 named; 277 eligible”).

Not named is not ranked last. A business with no cells naming it has a share of 0 and no rank. It is listed as “not named in this reading” on the ranking page and on its own entity page. The index never infers quality, safety, competence or compliance from a zero.

Engine score is the same calculation restricted to one engine's cells. Category and national figures are aggregates of the same cells; nothing is scored twice.

A change in weights, prompt set, engine list, run count or universe rule produces a new methodology version. Scores across versions are not compared on the same chart.

6. Publishing rules

Everything needed to check a score is public and free to use, with attribution, for checking, reporting, teaching and research. Commercial reuse or redistribution of the data files needs a data licence (see terms). Everything that could republish a false statement about a named business is withheld.

Published

• Ranking page per vertical × geography: every named business, share, spread, rank, engines, prompts named in, change, relationship disclosure; the count and list of eligible businesses not named.
• Entity page per eligible business, named or not, at a permanent URL. Each reading archived at /readings/{market}-{YYYY-MM}.
• Prompt page per prompt: which businesses each engine named, by reading.
• Cell file per reading (CSV and JSON): prompt id, engine, run, timestamp, businesses named with position, source URLs cited. No answer text.
• The prompt set, universe source and date, exclusions, methodology version and this document.

Withheld by default

• Full assistant answer text. Short factual excerpts may be quoted on an entity page where they name the business and contain no allegation.
• Any statement in an answer about a business's conduct, safety, complaints or legal standing, true or not. The index reports that a business was named, not what was said about it.
• Personal information about individuals, including named professionals, beyond the firm or practice name.

Cadence: readings publish within 10 days of the capture window closing. The next reading's window is shown on every page. A pipeline page lists verticals and geographies in capture and queued.

Technical: every table is server-rendered HTML; slugs never change; each page carries its capture date in the sitemap; data files are hosted at stable paths; crawlers for the scored engines and Google are permitted.

7. Governance and risk controls

The index describes itself in one way: it reports which eligible businesses AI assistants named in response to standardized prompts during a stated measurement window. It never says “the best businesses according to AI.”

Controls in force on every reading:

1. Preserve every prompt, response, model, account type, location setting, timestamp and citation.
2. Publish the eligible-universe size, exclusions, duplicate-resolution rule and minimum sample size.
3. Measure repeated runs and publish the spread, because answers are non-deterministic.
4. Call the metric “recommendation share” or “measured visibility”, never quality, trust, safety or best.
5. Never infer incompetence, danger or regulatory non-compliance from a zero or a fall.
6. Separate “not named” from “ranked last”; unnamed businesses tie at zero with no rank.
7. Do not publish entire assistant answers; publish structured cells and short factual excerpts.
8. Redact personal information, unsupported allegations and case-specific narratives.
9. Verify any regulatory or disciplinary fact against the regulator's own record before it appears on a page, and link it.
10. Run a correction and entity-merge process with a stated turnaround (section 10).
11. Require badge users to reproduce the date, geography, category and metric exactly (section 9).
12. Keep assistant-output evidence and buyer-behaviour evidence as separate claims; the index publishes the first and does not assert the second.
13. Name a responsible analyst on every reading; the publisher's founder does not sign readings in a vertical where a conflict is declared (section 8).

A reading that cannot meet controls 1–3 is not published. A reading published in breach of 4–8 is corrected on the page with a dated note.

8. Independence and conflicts

The Trust Index is published by LittPro Inc. LittPro Inc. and Hey Taylor, an AI-search agency, are related companies. Businesses that are Hey Taylor clients are marked 'Agency client'. That relationship is a conflict of interest, and this section is how it is handled.

No business can pay to be tested, listed, scored, moved or removed. Readings are unpaid. Eligibility comes from the universe source (section 2), scores from the published method (section 5). Hey Taylor staff have no access to the scoring pipeline and no role in universe construction, prompt design or publication decisions.

Client relationships are disclosed where the score sits. Any business that is a current or former (24 months) client of Hey Taylor or a Hey Taylor partner agency is marked on the ranking row and on its entity page. Client status does not change the score. The index does not state which services a client bought.

Licensed assessors. Agencies on the public register may run Trust Index assessments for individual businesses. An assessment is a private diagnostic for the business that commissions it. It never produces, changes or influences a Trust Index rating, ranking or badge. Only index readings produce ratings. Any assessor, Hey Taylor included, is licensed and listed only on the same published terms. An assessment of a business the assessor has also served in the previous 24 months is labelled as such.

Revenue the index earns is listed here: paid monitoring, available on identical terms to every eligible business whether named or not; badge licences, issued only for a result already published (section 9); data licensing of published files; assessor licences and practitioner certification, on published terms identical for every applicant. Badge revenue is not shared with sales partners per winner. No revenue line depends on any business's score.

Founder disclosure. The founder of Hey Taylor and the Trust Index serves on the board of Advent Health Care, an Ontario long-term care provider. In consequence:

• Long-term care homes are outside the scope of every Trust Index vertical. The retirement-homes vertical covers RHRA-licensed retirement homes only, which operate under a separate licensing regime.
• Entities affiliated with Advent Health Care are scored like every other eligible business, by the same pipeline and rules. The affiliation is disclosed on their listing, their profile and every reading that includes them.
• The founder takes no editorial, scoring or publication decision on the retirement-homes vertical. Each retirement-homes reading is signed by the named responsible analyst, and the recusal is noted on the reading.

Named analysts. Every reading names the analyst responsible for it. Analysts declare any interest in a ranked business before a reading; a declared interest means another analyst signs.

This section is updated whenever a relationship changes, and each change is dated in the version history.

9. Badges and permitted wording

A badge states a measured result, its date and its geography, and links to the reading it came from. It never says that AI ranked, chose, endorsed or trusts a business.

The Competition Act judges an advertisement by its general impression, not only its literal words, and a performance claim needs adequate substantiation. “Ranked #1 by AI assistants” implies the assistants conducted a ranking; they did not. The publisher ran prompts and computed a metric. The wording below states that.

Rules

• A badge is issued only for a published reading and expires when the next reading publishes. Expired badges must be removed or dated as historical.
• The licensee reproduces the metric, rank, geography, category and date exactly. Paraphrase voids the licence.
• Regulated professions use the listed form only until the relevant regulator's advertising rule has been reviewed for that vertical and the review is published here. Ontario lawyers: review against LSO Rule 4.2-1 pending. Ontario dentists and physicians: comparative claims are prohibited by RCDSO and CPSO advertising standards; the listed form is the ceiling.
• Retirement homes: no consumer-facing badge is issued in the first two readings. Operators may cite the reading with the full form and disclosure in materials aimed at the trade.
• US jurisdictions use state-specific wording. Florida permits bona fide ratings only where criteria are objective, uniform and verifiable; Texas requires advertising review. Badges for US firms are not issued until that state's variant is published.
• A badge licence never changes eligibility, score, copy or placement.

10. Corrections and entity merges

Any business, regulator or reader can request a correction from the link on every page. Requests are acknowledged within 2 business days and resolved within 10.

Every correction is logged with its date on the affected page and in the version history. A corrected reading keeps its original URL; the superseded figures stay visible as struck through for 90 days.

11. Version history

Position weights are stated in section 5 from v2.2 onward. Newest changes are listed first.

This methodology is published for transparency under CC BY-ND 4.0: share it unchanged, with credit. Modified versions may not be distributed. © 2026 LittPro Inc. Trust Index™ is a trademark of LittPro Inc. Trust Index ratings come only from index readings published by LittPro Inc. Only licensed assessors may run Trust Index assessments, and an assessment never produces a rating.